MGCBFormer: The multiscale grid-prior and class-inter boundary-aware transformer for polyp segmentation.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Deep learning for polyp segmentation aids in early colorectal cancer detection.
• Current supervised methods underutilize pixel-level labels and focus excessively on backbone networks.
• There's a need for methods that fully leverage existing polyp target information.

Purpose of the Study:

• To introduce the Multiscale Grid-prior and Class-inter boundary-aware Transformer (MGCBFormer) for enhanced polyp segmentation.
• To address limitations of mainstream supervised polyp segmentation techniques.
• To improve the efficiency and accuracy of polyp detection in intestinal imaging.

Main Methods:

• Development of MGCBFormer, incorporating a multiscale grid-prior and nested channel attention block (MGNAB).
• Integration of a class-inter boundary-aware block (CBB) with a boundary preprocessing strategy.
• Utilization of a global double-axis association coupler (GDAC) for deep supervision and noise filtering.

Main Results:

• MGCBFormer demonstrated superior predictive performance across five public polyp datasets (Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, CVC-300, ETIS-LaribPolypDB).
• Comparative experiments showed MGCBFormer outperformed twelve other polyp segmentation methods.
• The model exhibited strong generalization ability in polyp segmentation tasks.

Conclusions:

• MGCBFormer offers a significant advancement in deep learning-based polyp segmentation.
• The proposed architecture effectively mines polyp target information and boundary details.
• This technology holds promise for improving the speed and accuracy of colorectal cancer precursor diagnosis.